Video Analytics
Video analytics in SBN Media is built around the Image Classifier, which detects objects in video, together with a few supporting services that store, review, and cross-reference what it finds. This page explains those supporting services and which you need to run for which feature.
Overview
The Image Classifier does the detection (see its own page). Around it sit three supporting services:
- Track Snapshot - freezes each detection into a reviewable snapshot: the full frame, the bounding boxes, and a crop of each detected object.
- Annotation - lets operators label those snapshots (confirmed, false alarm, uncertain) so the detection models can be improved.
- Re-Identification - recognises the same person or vehicle across different cameras and over time, powering "find similar" searches.
Each runs one per host and is managed by the watchdog. None of them is needed simply to *detect* objects - they add review, labelling, and cross-camera matching on top.
What to run for which feature
| You want to... | Run |
|---|---|
| Detect objects in video | Image Classifier |
| Review detections (images, bounding boxes) | + Track Snapshot |
| Label detections to improve the models | + Annotation (which needs Track Snapshot) |
| Find the same person or vehicle across cameras | + Re-Identification (opt-in) |
The services
Track Snapshot
When the Image Classifier detects something, Track Snapshot freezes that moment into a permanent snapshot - the full frame, the bounding boxes, the best crop of each tracked object, and the settings the classifier used - and serves them back for review. It is what lets an operator see the evidence behind a detection.
It is strongly recommended wherever you want to see or keep a record of detections. Without it, detections are still raised, but there is no stored image to look at and the labelling and cross-camera features have nothing to work from. It has no settings of its own; snapshots are kept in SBN Media's own store until removed.
Annotation
Annotation lets an operator label a snapshot - confirmed, false alarm, or uncertain, with optional notes. Those labels are the feedback used to refine the detection models over time.
It is optional and only needed if you run a model-improvement programme. A label is attached to a snapshot, so Annotation depends on Track Snapshot. It has no settings of its own; labels are kept in SBN Media's own store.
Re-Identification
Re-Identification stores a compact "fingerprint" of each detected object so the same one can be recognised again - on another camera, or later in time. This is what powers cross-camera "find similar" searches.
It is optional and off by default. Turn it on with reidstore.enabled. It stores its fingerprints in a local SQLite database on the host. Fingerprints are only produced when the Image Classifier is configured with a re-identification model; without one, nothing is stored.
To produce the fingerprints, install a re-identification model alongside the detector on the Image Classifier host (see the Image Classifier page for the full setup):
# A generic embedder, suitable for most objects
./sbn-media classifier setup dinov2-small
# Better for people, on small surveillance crops
./sbn-media classifier setup solider-smallWith an embedder installed and reidstore.enabled set to true, the Image Classifier starts producing fingerprints and Re-Identification writes them to its store.
| Setting | Default | Description |
|---|---|---|
reidstore.enabled | false | Turn re-identification on. |
reidstore.retention.days | 7 | How many days fingerprints are kept before they are removed automatically. |
After changing these in sbn-media.local.yaml, apply them with ./sbn-media config reload.
FAQ
Do I need all of these just to detect objects?
No. Detection needs only the Image Classifier. Track Snapshot, Annotation, and Re-Identification add review, labelling, and cross-camera matching on top.
What is the minimum for operators to review detections?
Add Track Snapshot. It stores the frame and the detected objects so an operator can see what was detected; detections without it have no stored image.
What does Re-Identification need?
reidstore.enabled set to true and a re-identification model configured on the Image Classifier. Without the model, no fingerprints are produced.
Where is the data kept, and for how long?
Snapshots and labels are kept in SBN Media's own store until removed. Re-identification fingerprints are kept in the Re-Identification store for reidstore.retention.days (7 by default).
Related pages
- SBN Media Overview (
SBN-Media/overview) - Image Classifier (
SBN-Media/Media-Processing/image-classifier) - Installing and Configuring SBN Media (
SBN-Media/installation) - Watchdog (
SBN-Media/Platform/watchdog)